基于多任务学习的特征融合跨年龄人脸识别研究

Research on cross-age face recognition based on feature fusion with multi task learning

针对跨年龄人脸识别过程中面部纹理和形状随年龄衰老产生变化而使现有识别算法精度下降的问题,提出一种基于多任务学习的特征融合跨年龄人脸识别算法。首先提取人脸随年龄衰老较稳定的多尺度与多关键区域的局部二值特征(Local Binary Patterns,LBP),将其与人脸深度学习模型进行特征融合获得多层次的全局人脸特征,联合年龄有序预测任务与身份分类任务一起优化网络模型。算法在提取融合多层次的人脸特征后,通过年龄预测任务获得年龄特征,将年龄特征经全连接层函数映射为人脸衰老中发生改变的衰老特征,再在人脸全局特征中剔除衰老特征就可得到人脸衰老过程中不变的稳定特征。这种算法在公开的跨年龄人脸数据集上的识别精度较现有算法有了一定提升。

In the process of cross-age face id recognition,to address problems of decreased accuracy of existing recognition algorithms due to changes in the face texture and shape change with age,a feature fusion cross age face id recognition algorithm based on multi task learning is proposed.Firstly,we extract the local binary pattern of multi-scale and multi-key regions that are relatively stable with aging and integrate them with features of face id deep learning model to obtain multi-level global face id features.Then,the network model with the age-ordered prediction task and the identity classification task is optimized.After extracting and fusion the multi-level face id features,the algorithm obtains age features through age prediction tasks,The age features are mapped into aging features that change during facial aging through full-connectivity layer functions,and then eliminates the aging features from the global facial features to obtain the stable features that remain unchanged during the aging process.The recognition accuracy of the proposed algorithm on publicly available cross-age facial datasets has improved to some extent compared to existing algorithms.